thanhtvt
commited on
Commit
•
4ac7ffc
1
Parent(s):
bcd8e2f
remove alsd
Browse files
app.py
CHANGED
@@ -19,7 +19,7 @@ def get_duration(filename: str) -> float:
|
|
19 |
return librosa.get_duration(path=filename)
|
20 |
|
21 |
|
22 |
-
def
|
23 |
out_filename = os.path.splitext(in_filename)[0] + ".wav"
|
24 |
logging.info(f"Converting {in_filename} to {out_filename}")
|
25 |
y, sr = librosa.load(in_filename, sr=16000)
|
@@ -27,22 +27,6 @@ def convert_to_wav1(in_filename: str) -> str:
|
|
27 |
return out_filename
|
28 |
|
29 |
|
30 |
-
def convert_to_wav(in_filename: str) -> str:
|
31 |
-
"""Convert the input audio file to a wave file"""
|
32 |
-
out_filename = in_filename + ".wav"
|
33 |
-
logging.info(f"Converting '{in_filename}' to '{out_filename}'")
|
34 |
-
|
35 |
-
sp_args = ["ffmpeg", "-hide_banner", "-i", in_filename, "-ar", "16000", out_filename]
|
36 |
-
sp_args.insert(2, "-y") if os.path.exists(out_filename) else None
|
37 |
-
# Create a subprocess to run the ffmpeg command.
|
38 |
-
_ = subprocess.Popen(
|
39 |
-
sp_args,
|
40 |
-
stdin=subprocess.PIPE,
|
41 |
-
)
|
42 |
-
|
43 |
-
return out_filename
|
44 |
-
|
45 |
-
|
46 |
def build_html_output(s: str, style: str = "result_item_success"):
|
47 |
return f"""
|
48 |
<div class='result'>
|
@@ -58,7 +42,6 @@ def process_url(
|
|
58 |
decoding_method: str,
|
59 |
beam_size: int,
|
60 |
max_symbols_per_step: int,
|
61 |
-
max_out_seq_len_ratio: float,
|
62 |
):
|
63 |
logging.info(f"Processing URL: {url}")
|
64 |
with tempfile.NamedTemporaryFile() as f:
|
@@ -67,8 +50,7 @@ def process_url(
|
|
67 |
return process(in_filename=f.name,
|
68 |
decoding_method=decoding_method,
|
69 |
beam_size=beam_size,
|
70 |
-
max_symbols_per_step=max_symbols_per_step
|
71 |
-
max_out_seq_len_ratio=max_out_seq_len_ratio)
|
72 |
except Exception as e:
|
73 |
logging.info(str(e))
|
74 |
return "", build_html_output(str(e), "result_item_error")
|
@@ -79,7 +61,6 @@ def process_uploaded_file(
|
|
79 |
decoding_method: str,
|
80 |
beam_size: int,
|
81 |
max_symbols_per_step: int,
|
82 |
-
max_out_seq_len_ratio: float,
|
83 |
):
|
84 |
if in_filename is None or in_filename == "":
|
85 |
return "", build_html_output(
|
@@ -93,8 +74,7 @@ def process_uploaded_file(
|
|
93 |
return process(in_filename=in_filename,
|
94 |
decoding_method=decoding_method,
|
95 |
beam_size=beam_size,
|
96 |
-
max_symbols_per_step=max_symbols_per_step
|
97 |
-
max_out_seq_len_ratio=max_out_seq_len_ratio)
|
98 |
except Exception as e:
|
99 |
logging.info(str(e))
|
100 |
return "", build_html_output(str(e), "result_item_error")
|
@@ -105,7 +85,6 @@ def process_microphone(
|
|
105 |
decoding_method: str,
|
106 |
beam_size: int,
|
107 |
max_symbols_per_step: int,
|
108 |
-
max_out_seq_len_ratio: float,
|
109 |
):
|
110 |
if in_filename is None or in_filename == "":
|
111 |
return "", build_html_output(
|
@@ -119,8 +98,7 @@ def process_microphone(
|
|
119 |
return process(in_filename=in_filename,
|
120 |
decoding_method=decoding_method,
|
121 |
beam_size=beam_size,
|
122 |
-
max_symbols_per_step=max_symbols_per_step
|
123 |
-
max_out_seq_len_ratio=max_out_seq_len_ratio)
|
124 |
except Exception as e:
|
125 |
logging.info(str(e))
|
126 |
return "", build_html_output(str(e), "result_item_error")
|
@@ -131,7 +109,6 @@ def process(
|
|
131 |
decoding_method: str,
|
132 |
beam_size: int,
|
133 |
max_symbols_per_step: int,
|
134 |
-
max_out_seq_len_ratio: float,
|
135 |
):
|
136 |
logging.info(f"in_filename: {in_filename}")
|
137 |
|
@@ -148,8 +125,7 @@ def process(
|
|
148 |
recognizer = UETASRModel(repo_id,
|
149 |
decoding_method,
|
150 |
beam_size,
|
151 |
-
max_symbols_per_step
|
152 |
-
max_out_seq_len_ratio)
|
153 |
text = recognizer.predict(filename)
|
154 |
|
155 |
date_time = now.strftime("%d/%m/%Y, %H:%M:%S.%f")
|
@@ -167,7 +143,7 @@ def process(
|
|
167 |
"""
|
168 |
if rtf > 1:
|
169 |
info += (
|
170 |
-
"<br/>We are loading
|
171 |
"Please run again to measure the real RTF.<br/>"
|
172 |
)
|
173 |
|
@@ -202,59 +178,40 @@ with demo:
|
|
202 |
|
203 |
decode_method_radio = gr.Radio(
|
204 |
label="Decoding method",
|
205 |
-
choices=["greedy_search", "beam_search"
|
206 |
value="greedy_search",
|
207 |
interactive=True,
|
208 |
)
|
209 |
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
)
|
219 |
|
220 |
-
def
|
221 |
-
if decoding_method
|
222 |
-
return gr.update(
|
223 |
else:
|
224 |
-
return gr.update(
|
225 |
|
226 |
-
decode_method_radio.change(
|
|
|
|
|
227 |
|
228 |
max_symbols_per_step_slider = gr.Slider(
|
229 |
label="Maximum symbols per step",
|
230 |
minimum=1,
|
231 |
-
maximum=
|
232 |
step=1,
|
233 |
value=5,
|
234 |
interactive=True,
|
235 |
visible=True,
|
236 |
)
|
237 |
|
238 |
-
max_out_seq_len_slider = gr.Slider(
|
239 |
-
label="Maximum output sequence length ratio",
|
240 |
-
minimum=0,
|
241 |
-
maximum=1,
|
242 |
-
step=0.01,
|
243 |
-
value=0.6,
|
244 |
-
interactive=True,
|
245 |
-
visible=False,
|
246 |
-
)
|
247 |
-
|
248 |
-
def switch_slider(decoding_method):
|
249 |
-
if decoding_method == "alsd_search":
|
250 |
-
return gr.update(visible=False), gr.update(visible=True)
|
251 |
-
else:
|
252 |
-
return gr.update(visible=True), gr.update(visible=False)
|
253 |
-
|
254 |
-
decode_method_radio.change(switch_slider,
|
255 |
-
decode_method_radio,
|
256 |
-
[max_symbols_per_step_slider, max_out_seq_len_slider])
|
257 |
-
|
258 |
with gr.Tabs():
|
259 |
with gr.TabItem("Upload from disk"):
|
260 |
uploaded_file = gr.Audio(
|
@@ -308,9 +265,8 @@ with demo:
|
|
308 |
inputs=[
|
309 |
uploaded_file,
|
310 |
decode_method_radio,
|
311 |
-
|
312 |
max_symbols_per_step_slider,
|
313 |
-
max_out_seq_len_slider,
|
314 |
],
|
315 |
outputs=[uploaded_output, uploaded_html_info],
|
316 |
)
|
@@ -320,9 +276,8 @@ with demo:
|
|
320 |
inputs=[
|
321 |
microphone,
|
322 |
decode_method_radio,
|
323 |
-
|
324 |
max_symbols_per_step_slider,
|
325 |
-
max_out_seq_len_slider,
|
326 |
],
|
327 |
outputs=[recorded_output, recorded_html_info],
|
328 |
)
|
@@ -332,9 +287,8 @@ with demo:
|
|
332 |
inputs=[
|
333 |
url_textbox,
|
334 |
decode_method_radio,
|
335 |
-
|
336 |
max_symbols_per_step_slider,
|
337 |
-
max_out_seq_len_slider,
|
338 |
],
|
339 |
outputs=[url_output, url_html_info],
|
340 |
)
|
|
|
19 |
return librosa.get_duration(path=filename)
|
20 |
|
21 |
|
22 |
+
def convert_to_wav(in_filename: str) -> str:
|
23 |
out_filename = os.path.splitext(in_filename)[0] + ".wav"
|
24 |
logging.info(f"Converting {in_filename} to {out_filename}")
|
25 |
y, sr = librosa.load(in_filename, sr=16000)
|
|
|
27 |
return out_filename
|
28 |
|
29 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
def build_html_output(s: str, style: str = "result_item_success"):
|
31 |
return f"""
|
32 |
<div class='result'>
|
|
|
42 |
decoding_method: str,
|
43 |
beam_size: int,
|
44 |
max_symbols_per_step: int,
|
|
|
45 |
):
|
46 |
logging.info(f"Processing URL: {url}")
|
47 |
with tempfile.NamedTemporaryFile() as f:
|
|
|
50 |
return process(in_filename=f.name,
|
51 |
decoding_method=decoding_method,
|
52 |
beam_size=beam_size,
|
53 |
+
max_symbols_per_step=max_symbols_per_step)
|
|
|
54 |
except Exception as e:
|
55 |
logging.info(str(e))
|
56 |
return "", build_html_output(str(e), "result_item_error")
|
|
|
61 |
decoding_method: str,
|
62 |
beam_size: int,
|
63 |
max_symbols_per_step: int,
|
|
|
64 |
):
|
65 |
if in_filename is None or in_filename == "":
|
66 |
return "", build_html_output(
|
|
|
74 |
return process(in_filename=in_filename,
|
75 |
decoding_method=decoding_method,
|
76 |
beam_size=beam_size,
|
77 |
+
max_symbols_per_step=max_symbols_per_step)
|
|
|
78 |
except Exception as e:
|
79 |
logging.info(str(e))
|
80 |
return "", build_html_output(str(e), "result_item_error")
|
|
|
85 |
decoding_method: str,
|
86 |
beam_size: int,
|
87 |
max_symbols_per_step: int,
|
|
|
88 |
):
|
89 |
if in_filename is None or in_filename == "":
|
90 |
return "", build_html_output(
|
|
|
98 |
return process(in_filename=in_filename,
|
99 |
decoding_method=decoding_method,
|
100 |
beam_size=beam_size,
|
101 |
+
max_symbols_per_step=max_symbols_per_step)
|
|
|
102 |
except Exception as e:
|
103 |
logging.info(str(e))
|
104 |
return "", build_html_output(str(e), "result_item_error")
|
|
|
109 |
decoding_method: str,
|
110 |
beam_size: int,
|
111 |
max_symbols_per_step: int,
|
|
|
112 |
):
|
113 |
logging.info(f"in_filename: {in_filename}")
|
114 |
|
|
|
125 |
recognizer = UETASRModel(repo_id,
|
126 |
decoding_method,
|
127 |
beam_size,
|
128 |
+
max_symbols_per_step)
|
|
|
129 |
text = recognizer.predict(filename)
|
130 |
|
131 |
date_time = now.strftime("%d/%m/%Y, %H:%M:%S.%f")
|
|
|
143 |
"""
|
144 |
if rtf > 1:
|
145 |
info += (
|
146 |
+
"<br/>We are loading required resources for the first run. "
|
147 |
"Please run again to measure the real RTF.<br/>"
|
148 |
)
|
149 |
|
|
|
178 |
|
179 |
decode_method_radio = gr.Radio(
|
180 |
label="Decoding method",
|
181 |
+
choices=["greedy_search", "beam_search"],
|
182 |
value="greedy_search",
|
183 |
interactive=True,
|
184 |
)
|
185 |
|
186 |
+
beam_size_slider = gr.Slider(
|
187 |
+
label="Beam size",
|
188 |
+
minimum=1,
|
189 |
+
maximum=20,
|
190 |
+
step=1,
|
191 |
+
value=1,
|
192 |
+
interactive=False,
|
193 |
+
)
|
|
|
194 |
|
195 |
+
def interact_beam_slider(decoding_method):
|
196 |
+
if decoding_method == "greedy_search":
|
197 |
+
return gr.update(value=1, interactive=False)
|
198 |
else:
|
199 |
+
return gr.update(interactive=True)
|
200 |
|
201 |
+
decode_method_radio.change(interact_beam_slider,
|
202 |
+
decode_method_radio,
|
203 |
+
beam_size_slider)
|
204 |
|
205 |
max_symbols_per_step_slider = gr.Slider(
|
206 |
label="Maximum symbols per step",
|
207 |
minimum=1,
|
208 |
+
maximum=20,
|
209 |
step=1,
|
210 |
value=5,
|
211 |
interactive=True,
|
212 |
visible=True,
|
213 |
)
|
214 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
215 |
with gr.Tabs():
|
216 |
with gr.TabItem("Upload from disk"):
|
217 |
uploaded_file = gr.Audio(
|
|
|
265 |
inputs=[
|
266 |
uploaded_file,
|
267 |
decode_method_radio,
|
268 |
+
beam_size_slider,
|
269 |
max_symbols_per_step_slider,
|
|
|
270 |
],
|
271 |
outputs=[uploaded_output, uploaded_html_info],
|
272 |
)
|
|
|
276 |
inputs=[
|
277 |
microphone,
|
278 |
decode_method_radio,
|
279 |
+
beam_size_slider,
|
280 |
max_symbols_per_step_slider,
|
|
|
281 |
],
|
282 |
outputs=[recorded_output, recorded_html_info],
|
283 |
)
|
|
|
287 |
inputs=[
|
288 |
url_textbox,
|
289 |
decode_method_radio,
|
290 |
+
beam_size_slider,
|
291 |
max_symbols_per_step_slider,
|
|
|
292 |
],
|
293 |
outputs=[url_output, url_html_info],
|
294 |
)
|
decode.py
CHANGED
@@ -1,7 +1,7 @@
|
|
1 |
import logging
|
2 |
import tensorflow as tf
|
3 |
from functools import lru_cache
|
4 |
-
from uetasr.searchers import GreedyRNNT, BeamRNNT
|
5 |
|
6 |
|
7 |
@lru_cache(maxsize=5)
|
@@ -12,7 +12,6 @@ def get_searcher(
|
|
12 |
text_decoder: tf.keras.layers.experimental.preprocessing.PreprocessingLayer,
|
13 |
beam_size: int,
|
14 |
max_symbols_per_step: int,
|
15 |
-
max_output_seq_length_ratio: float,
|
16 |
):
|
17 |
common_kwargs = {
|
18 |
"decoder": decoder,
|
@@ -32,12 +31,6 @@ def get_searcher(
|
|
32 |
alpha=0.0,
|
33 |
**common_kwargs,
|
34 |
)
|
35 |
-
elif searcher_type == "alsd_search":
|
36 |
-
searcher = ALSDBeamRNNT(
|
37 |
-
fraction=max_output_seq_length_ratio,
|
38 |
-
beam_size=beam_size,
|
39 |
-
**common_kwargs,
|
40 |
-
)
|
41 |
else:
|
42 |
logging.info(f"Unknown searcher type: {searcher_type}")
|
43 |
|
|
|
1 |
import logging
|
2 |
import tensorflow as tf
|
3 |
from functools import lru_cache
|
4 |
+
from uetasr.searchers import GreedyRNNT, BeamRNNT
|
5 |
|
6 |
|
7 |
@lru_cache(maxsize=5)
|
|
|
12 |
text_decoder: tf.keras.layers.experimental.preprocessing.PreprocessingLayer,
|
13 |
beam_size: int,
|
14 |
max_symbols_per_step: int,
|
|
|
15 |
):
|
16 |
common_kwargs = {
|
17 |
"decoder": decoder,
|
|
|
31 |
alpha=0.0,
|
32 |
**common_kwargs,
|
33 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
else:
|
35 |
logging.info(f"Unknown searcher type: {searcher_type}")
|
36 |
|
model.py
CHANGED
@@ -101,7 +101,6 @@ class UETASRModel:
|
|
101 |
decoding_method: str,
|
102 |
beam_size: int,
|
103 |
max_symbols_per_step: int,
|
104 |
-
max_output_seq_length_ratio: float,
|
105 |
):
|
106 |
self.featurizer, self.encoder_model, jointer, decoder, text_encoder, self.model = _get_conformer_pre_trained_model(repo_id)
|
107 |
self.searcher = get_searcher(
|
@@ -111,7 +110,6 @@ class UETASRModel:
|
|
111 |
text_encoder,
|
112 |
beam_size,
|
113 |
max_symbols_per_step,
|
114 |
-
max_output_seq_length_ratio,
|
115 |
)
|
116 |
|
117 |
def predict(self, in_filename: str):
|
|
|
101 |
decoding_method: str,
|
102 |
beam_size: int,
|
103 |
max_symbols_per_step: int,
|
|
|
104 |
):
|
105 |
self.featurizer, self.encoder_model, jointer, decoder, text_encoder, self.model = _get_conformer_pre_trained_model(repo_id)
|
106 |
self.searcher = get_searcher(
|
|
|
110 |
text_encoder,
|
111 |
beam_size,
|
112 |
max_symbols_per_step,
|
|
|
113 |
)
|
114 |
|
115 |
def predict(self, in_filename: str):
|
requirements.txt
CHANGED
@@ -1,2 +1,2 @@
|
|
1 |
-
uetasr @ git+https://github.com/thanhtvt/uetasr
|
2 |
requests==2.28.2
|
|
|
1 |
+
uetasr @ git+https://github.com/thanhtvt/uetasr@v0.2.1
|
2 |
requests==2.28.2
|